Open Models Almanac
Methodsinference & servinginference quantization

specific method Β· filed under inference & serving

Four-Over-Six

Four-Over-Six is an FP4 weight scale-selection rule for post-training quantization that sets the per-block NVFP4 weight scale.1,2 For each weight microblock it chooses between the M = 4 and M = 6 FP4 grids, picking whichever option minimizes reconstruction error.3

Trade-off It trades a small amount of additional zero-rounding for better handling of high-magnitude tails.4

How this definition was written

Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.

  1. Four-Over-Six (Cook et al.) scaling β€” NVIDIA Nemotron 3 Ultra Technical Report
  2. Each algorithm changes the per-block NVFP4 scale β€” NVIDIA Nemotron 3 Ultra Technical Report
  3. Four-Over-Six selects between the𝑀 = 4and 𝑀 = 6weight grids per block, using the option that minimizes β€” NVIDIA Nemotron 3 Ultra Technical Report
  4. This trades a small amount of additional zero-rounding for better handling of high-magnitude tails β€” NVIDIA Nemotron 3 Ultra Technical Report

Also called Four-Over-Six FP4 weight scale selection, Four-Over-Six scale selection method, Four-Over-Six weight scaling.

sources
2
model
1
lab adopt it
1
strongest
used

Documented in

Evidence

3 quotes from 2 sources, one card per model, each opening with its strongest.

How sources treat it
used 3

NVIDIA-Nemotron-3-Ultra-550B-A55B

NVIDIAadopted2 sources

Four-Over-Six selects between the 𝑀 = 4 and 𝑀 = 6 weight grids per block, using the option that minimizes reconstruction error.

usedpost trainingin Nemotron 3 UltraNVIDIA
2 more from this model

we selected Four-Over-Six for setting the FP4 routed-expert weight scales

usedpost trainingin Nemotron 3 UltraNVIDIA

we selected Four-Over-Six for setting the FP4 routed-expert weight scales

usedpost trainingin Nemotron 3 UltraNVIDIA
↑